Methodology & Caveats

Data sources

Kaggle Steam dataset (backfill) · Steam storefront API (metadata) · store-page HTML (AI Generated Content Disclosure — not exposed by any API) · Steam review histograms (first-month counts) · SteamSpy (owners ranges, playtime) · Gamalytic (sales estimates). Collection is keyless and rate-limited.

Reviews as a sales proxy

True sales are visible only to each developer. Like all public Steam research, we use review counts as the sales proxy (the "Boxleiter method"); first-month reviews are the outcome, matching Game Oracle's published design so results are directly comparable.

Model

Negative-binomial regression of first-month reviews on AI disclosure, controlling for genre, release month, launch-price band, publisher backing, and developer prior releases. Cohort: paid games released January 2024 onward.

Limitations

• Disclosure compliance is assumed; undisclosed AI use biases effects toward zero.
• Observational estimates, not proof of causation; unmeasured confounders (marketing spend, team talent) remain possible.
• First-month counts from month-granularity histograms are approximate.
• Owners/sales estimates carry wide error bars by nature.